RS-Prune: Efficient Data Pruning for Remote Sensing Diffusion Models
Analysis
This paper addresses the challenge of training efficient remote sensing diffusion models by proposing a training-free data pruning method called RS-Prune. The method aims to reduce data redundancy, noise, and class imbalance in large remote sensing datasets, which can hinder training efficiency and convergence. The paper's significance lies in its novel two-stage approach that considers both local information content and global scene-level diversity, enabling high pruning ratios while preserving data quality and improving downstream task performance. The training-free nature of the method is a key advantage, allowing for faster model development and deployment.
Key Takeaways
- •Proposes a training-free data pruning method (RS-Prune) for remote sensing diffusion models.
- •RS-Prune uses a two-stage approach considering local information and global scene diversity.
- •Achieves high pruning ratios (e.g., 85%) while improving convergence and generation quality.
- •Demonstrates state-of-the-art performance on downstream tasks like super-resolution and semantic image synthesis.
“The method significantly improves convergence and generation quality even after pruning 85% of the training data, and achieves state-of-the-art performance across downstream tasks.”